Burn Severity Drivers in Italian Large Wildﬁres

: The increase of wildﬁre incidence in highly populated areas signiﬁcantly enhances the risk for ecosystems and human lives, activities and infrastructures. In central and southern Italy, recent decades’ ﬁre records indicate that 2007 and 2017 were extreme years in terms of the number of ﬁres and total burned area. Among them, we selected large ﬁre events and explored their features and drivers of burn severity. We used a standardized extraction procedure to identify large wildﬁres (>100 ha) from the MODIS burned areas database and Landsat multi-spectral images. We mapped burn severity with the Relative Difference Normalized Burn Ratio index and explored the main drivers of severity using topographic, land-cover and anthropogenic predictors. We selected 113 wildﬁres for a collective total burned area of over 100,000 ha. Large ﬁres were more frequent in the southern than in the central and northern regions, especially in July and August. The average ﬁre size was about 900 ha and occurred mainly in shrublands (30.4%) and broadleaf forests (19.5%). With a random forest model, we observed that the highest severity occurred in conifer plantations and shrublands, in highly populated areas and at lower elevations. Burn severity models, at the landscape or regional scales, can be very useful tools for pre-and post-ﬁre forest management planning.


Introduction
Forest fire regimes have changed throughout history in many landscapes worldwide [1], especially in relation to shifts in anthropogenic activities and climate [2]. Humans have played a major role in shaping landscapes and modifying fuel loads and continuity [1]. Woody fuel density increases observed in many developed countries in recent decades are often the result of fire suppression and reforestation actions that have altered fire regimes [3]. Historically, the Mediterranean area hosts some of the largest and most severe wildfires in Europe, making up 90% of the total European burned area [4]. Since at least the Miocene, fire has been an integral component of Mediterranean ecosystems, and there is evidence that rising global temperatures will alter fire regimes [5]. Recent climate changes and land use/land cover changes in the Mediterranean basin are expected to be synchronous drivers of fire-regime shifts [6,7]. Temperature anomalies and summer droughts [8] are critical in explaining fire occurrences but, especially in the Mediterranean basin, land-use/land cover changes are additional drivers of raising wildfire frequency, size and severity [7,[9][10][11]. The increase in wildfire incidence in such highly populated areas introduces a high risk to human lives, activities and infrastructure. Nearly 200 million people live in just five countries: Greece, Italy, France, Spain and Portugal [12], all featuring an extensive wildland-urban interface (WUI), where the risk of fire spread is exceptionally high. In these countries, the rapid industrialization of the 20th century caused relevant policies and prevention, the five-year average burned area in Italy decreased from 147,150 ha in 1980-1989 to 72,945 ha in 2010-2017. Future simulated scenarios predict that Italy, together with other European countries, would experience a significant increase in extreme climatic events [45], likely affecting wildfire frequency and severity. Fire risk forecast and fire control planning are challenging without consistent records and complete time series at the regional or national level [46]. In Italy, the national wildfire inventory is a collection of non-standardized regional records, which is often visually assessed in the field with Global Navigation Satellite System (GNSS) devices after the fire occurrence and lacks the necessary consistency for suitable spatial distribution and severity analysis. A very recent study reports that the most harmonized data are related to the 2004-2012 interval, but these data are not consistent for all the administrative regions nor easily accessible [47].
The landscape simplification, triggered by the overall abandonment of marginal rural areas and forest/shrub expansion that started in the 1950-1960s [13][14][15], together with climate warming and severe drought events, increased the vulnerability of the Italian peninsula to wildfire ignition, occurrence and severity.
The aims of this work are (i) to map a sample of large-fires in Italy and compute burn severity within their perimeters with Landsat-derived images; (ii) to describe the environmental conditions (geographic distribution, topography and land-cover type) of the selected wildfires; and (iii) to model burn severity using homogenous broad-scale biophysical and anthropogenic predictors. We applied a standardized selection procedure of large forest fires that occurred in peninsular and insular Italy in 2007 and 2017, both years with a high number of fires and large burned areas due to extreme fire weather. We considered "large" a wildfire causing a burned area ≥100 hectares, the minimum area used in the literature concerning Mediterranean wildfires (e.g., Mancini et al. 2017 [48]). We hypothesized that pre-fire land cover and anthropogenic factors could be the most influential drivers for fire burn severity in the Mediterranean context. Investigation of burn severity is a crucial aspect for limiting the devastating effects of large forest fires on human lives, infrastructures and environmental resources; and statistical models could be very useful tools for pre-and post-fire forest management planning.

The Study Area
The analysis covers 209,198 km 2 , 69% of the Italian land surface, extending between 44 • and 37 • N and 8 • to 17 • E, from Liguria to Calabria and including the two largest islands (Sicily and Sardinia) ( Figure 1). In this area, wildfires occur mainly in the summer months, unlike in the Alpine regions. In the past few decades, socio-economic changes have caused a diffuse abandonment of rural activities and settlements in inland areas and new forests have naturally and extensively expanded in agro-pastoral lands, reducing the landscape heterogeneity and increasing fire risk [13].
Most of the study area is rugged, with 28.7% of the area defined as mountains (areas ≥ 600 m a.s.l.) and 50.3% as hills (≤600 m a.s.l.) [49]. The most prominent substrata are terrigenous, clastic and carbonate, with interspersed igneous and metamorphic outcrops. Climatically, the area is characterized by both the temperate (mean annual temperatures of around 10 • C and annual precipitation peaks in autumn and spring) and the Mediterranean climate (mean annual temperatures of around 13 • C and summer drought of 2-3 months) [50]. North-eastern slopes (Adriatic side) are, in general, more continental and have higher precipitation than south-western ones (Tyrrhenian side) [15]. Forests, on average, cover over 35% of the national land surface, but in some regions, they surpass 50% (e.g., Liguria, Tuscany, Umbria and Sardinia) [51]. Broadleaf forests occupy approximately 68% of the total forest cover, followed by conifer forests (13.5%) and mixed forests (9.7%) [51]. Fires are more frequent and larger in the southern and insular regions (Sicily, Sardinia, Calabria and Campania), with peaks in July and August [27]. Most forest fires occur in mountain areas and are primarily human-induced (over 75% in 1997-2007) [52]. We extracted a large-wildfire dataset, including two exceptional years, 2007 and 2017, when a large number of wildfires occurred. In these years, 227,730 ha and 161,987 ha were burned, respectively [44]. The high incidence of wildfire occurrences and total area burned has been driven by the annual climatic anomalies: in 2007, total precipitation was 16% lower and mean temperatures were 1.2 • C higher than average [53]; in 2017, precipitation was 22% lower and temperatures were 1.3 • C higher than average [54]. Most of the study area is rugged, with 28.7% of the area defined as mountains (areas ≥600 m a.s.l.) and 50.3% as hills (≤600 m a.s.l.) [49]. The most prominent substrata are terrigenous, clastic and carbonate, with interspersed igneous and metamorphic outcrops. Climatically, the area is characterized by both the temperate (mean annual temperatures of around 10° C and annual precipitation peaks in autumn and spring) and the Mediterranean climate (mean annual temperatures of around 13 °C and summer drought of 2-3 months) [50]. North-eastern slopes (Adriatic side) are, in general, more continental and have higher precipitation than south-western ones (Tyrrhenian side) [15]. Forests, on average, cover over 35% of the national land surface, but in some regions, they surpass 50% (e.g., Liguria, Tuscany, Umbria and Sardinia) [51]. Broadleaf forests occupy approximately 68% of the total forest cover, followed by conifer forests (13.5%) and mixed forests (9.7%) [51]. Fires are more frequent and larger in the southern and insular regions (Sicily, Sardinia, Calabria and Campania), with peaks in July and August [27]. Most forest fires occur in mountain areas and are primarily human-induced (over 75% in 1997-2007) [52]. We extracted a large-wildfire dataset, including two exceptional years, 2007 and 2017, when a large number of wildfires occurred. In these years, 227,730 ha and 161,987 ha were burned, respectively [44]. The high incidence of wildfire occurrences and total area burned has been driven by the annual climatic anomalies: in 2007, total precipitation was 16% lower and mean temperatures were 1.2 °C higher than average [53]; in 2017, precipitation was 22% lower and temperatures were 1.3 °C higher than average [54].

Wildfire Selection and Burn Severity Assessment
We set up a standardized procedure applied to the entire study area for the fire selection and further processing ( Figure 2). We searched for large wildfires (≥100 ha) using MODIS Collection 6 Level 3 (monthly burned area products, MCD64A1, 500 m of spatial resolution) [31]. In QGIS 3.4.0 software, we first filtered the MODIS dataset to keep only

Wildfire Selection and Burn Severity Assessment
We set up a standardized procedure applied to the entire study area for the fire selection and further processing ( Figure 2). We searched for large wildfires (≥100 ha) using MODIS Collection 6 Level 3 (monthly burned area products, MCD64A1, 500 m of spatial resolution) [31]. In QGIS 3.4.0 software, we first filtered the MODIS dataset to keep only burned areas larger than five pixels (about 125 ha). We used the "BurnDate" information derived from the MODIS data product of burned areas to search for Landsat scenes from the United States Geological Survey (USGS) Earth Resources Observation and Science (EROS) Center Science Processing Architecture (ESPA) On-Demand Interface. We collected pre-and post-fire images with the least cloud-contaminated acquisitions (<10% of cloud cover) available during the vegetation growing season. We found 59 Landsat TM and OLI images (30 m spatial resolution) covering the entire study area (Supplementary Materials Table S1).
The latest available pre-fire images before each fire event were used. Then, we used images from the following year to detect the fire scar. Landsat 5 TM-surface reflectance products were generated by the USGS using the Landsat Ecosystem Disturbance Adaptive Processing System (LEDAPS) [55], whereas Landsat 8 OLI image processing was based on the Landsat Surface Reflectance Code (LaSRC) [56]. We masked clouds, cloud shadows and water bodies with the Fmask 3.3 tool in each acquisition [57]. We assessed the fire perimeter and the spatial distribution of burn severity using the Relative Difference Normalized Burn Ratio (RdNBR) [40].  Table S1).

Figure 2.
Flowchart of the standardized procedure for wildfire selection, perimeter extraction and burn severity assessment: input data sources (yellow), filtering procedure (green), software used (blue) and output map (pink).
The latest available pre-fire images before each fire event were used. Then, we used images from the following year to detect the fire scar. Landsat 5 TM-surface reflectance products were generated by the USGS using the Landsat Ecosystem Disturbance Adaptive Processing System (LEDAPS) [55], whereas Landsat 8 OLI image processing was based on the Landsat Surface Reflectance Code (LaSRC) [56]. We masked clouds, cloud shadows and water bodies with the Fmask 3.3 tool in each acquisition [57]. We assessed the fire perimeter and the spatial distribution of burn severity using the Relative Difference Normalized Burn Ratio (RdNBR) [40]. Figure 2. Flowchart of the standardized procedure for wildfire selection, perimeter extraction and burn severity assessment: input data sources (yellow), filtering procedure (green), software used (blue) and output map (pink).
We detected spectral changes one year after the fire as an extended assessment of remotely sensed burn severity, including first-and second-order effects caused by fire [21]. For the severity assessment, we calibrated the dNBR for single wildfires with a dNBR offset; specifically, we averaged dNBR values of pixels falling within two/three undisturbed polygons closest to the fire of the same land-cover types that mainly burned in the event. This procedure minimizes non-fire-induced reflectance variation associated with inter-annual plant phenology and precipitation variability [40,42]. We adopted a semiautomatic method for fire scar segmentation and perimeter extraction. We applied automatic segmentation using eCognition Developer 64 software (scale factor = 30, color factor = 0.5) and a manual on-screen extraction [21] of burned segmented polygons. Then, we corrected perimeters through on-screen digitization of burned pixels in post-fire Landsat TM/OLI images using false-color composites (RGB = SWIR2, NIR, Red), since misclassification between unchanged and low-severity pixels can frequently occur [36].
Finally, through filtering, we selected the wildfires larger than 100 ha (following Mancini et al. 2017 [48]) and with at least 50% of vegetated land cover types within the perimeters, using CORINE Land Cover (CLC, minimum mapping unit of 25 ha) 2006 (CLC06, [58]) for 2007 and CORINE Land Cover 2012 (CLC12, [59]) for 2017 fires. The filtering retained around 25% of the initial MCD64 burned area records and included 113 selected wildfires that we used in the analysis. For each selected fire, we overlapped the MODIS burned area, the Landsat fire perimeter and the fire severity maps (Figure 3).  Table S2 of Supplementary Materials for more information on the selected fires.

Dataset Creation and Feature Description
We extracted burned surface area (ha), perimeter length (km) and regions of fire occurrence from each fire polygon by working with a homogeneous projected coordinate system (WGS84UTM33N). We attributed each location to the most-represented municipality within each fire perimeter and assigned an identification code (ID) to each event detected (Supplementary Materials Table S2). We used a Digital Elevation Model 20 × 20 m (DEM SINAnet ISPRA) to determine topographic features of the fire surface area: we extracted in the GIS environment elevation (ELE), slope (SLO), Heat Load Index (HLI) as a proxy of slope exposure [60] and Topographic Wetness Index (TWI) as a proxy of terrain wetness [61]. We used CLC06 (2007 fires) and CLC12 (2017 fires) (Level 3-4-5) to check the land-cover types within fire perimeters [62]. We merged the original CLC categories into eleven larger groups: urban areas (URB), water bodies (WTB), orchards (ORC), croplands (CRP), heterogeneous areas with agriculture and forest (AGF), shrublands (SHB), sparse vegetation (SVG), herbaceous vegetation and grasslands (GRS), broadleaf forests (BRF), conifer forests (COF) and mixed forests (MXF) ( Table 1).

Severity Modelling
We explored the main drivers of burn severity by modelling within-fire RdNBR using topographic features (ELE, SLO, HLI and TWI) and nine land-cover types (CRP, ORC, AGF, SHB, SVG, GRS, BRF, COF and MXF) as predictors. We excluded urban areas and waterbodies from the analysis. We also included three anthropogenic variables: (i) the Euclidean distance from state and province roads (ROA) extracted from the National Geoportal database [63]; (ii) the Euclidean distance from railroads (RAI) extracted from OpenStreetMap; and (iii) the population density of municipalities that fall within the fire footprint (POP) from the national population census 2001 for 2007 fires and 2011 for 2017 fires [64,65] (Table 1). We used per-pixel RdNBR as the dependent variable, ensuring a comparison of burn severity between different vegetation covers, e.g., shrublands and forests, and pre-fire vegetation densities [40]. We processed all records in a GIS environment, and we spatially joined population densities with their associated municipality vector maps and finally rasterized. We rescaled the spatial resolution of the whole dataset to the Landsat pixel (30 m), and we clipped raster layers with fire perimeters.
We used a large sample of over one million observations from a 30 m point grid where single points correspond to Landsat pixel centroids. We attributed descriptive information (fire ID, fire YEAR), the dependent variable (RdNBR) and predictor values to each point in a GIS environment. Then, we calculated Pearson's correlation matrices of independent quantitative variables to check for multi-collinearity (R < 0.6). To limit the influence of spatial autocorrelation, we subsampled observations using a minimum distance of 250 m prior to fitting a Random Forest (RF) regression model [66].The subsampling approach reduced the size of the dataset to 12,961 observations. We checked the degree of intensity of spatial autocorrelation between the residuals of the RF model at distance steps of 250 m, from 250 to 6000 m, through a semivariogram ( Figure S1) computed with the "geoR" R package [67].
The RF model can assess the relationship between Landsat-derived burn severity and several environmental and human-related drivers (Table 1). We employed the RF implementation in the "ranger" [68] R package [69] and optimized hyper-parameters (num.trees = 500; mtry = 1; sample.fraction = 0.2; min.node.size = 10) through the "mlrMBO" R package [70]. To assess the performance of the model, we performed a five-fold spatial cross-validation repeated 10 times and computed the average root mean square error (RMSE). Spatial partitions were based on k-means clustering of observation coordinates and allowed us to maintain independence between training and validation sets [71]. We assessed variable importance through the permutation method [66] and employed partial dependencies [72,73] to interpret the marginal effects of individual predictors on Landsatderived burn severity.

Results
Out of the 113 large wildfires used for the analysis, 49 occurred in 2007 (burned area 56,765 ha) and 64 in 2017 (45,800 ha), for a total burned area of 102,565 ha ( Figure 4; Table 2). The mean burned area was 908 ha (±1237) with a range of 108-7474 ha. The largest fire (7474 ha) was recorded in 2007 in Sardinia; the smallest one (108 ha) in Calabria in 2017 (Supplementary Materials Table S2). Most wildfires were in southern Italy (about 75% of fires and 73% of the total burned area), especially in Calabria (n = 34), Sicily (n = 22), Campania (n = 16) and Sardinia (n = 13). Sicily was the most affected region, with a total of 30,769 ha burned in the two years ( Table 2).  About one-third (36%) of burned areas had an extent ≤500 ha, while those ranging from 250 to 750 ha were about 50% ( Figure 5). They mainly occurred in the summer, in July (50%) and in August (42%). These fires occurred within a wide elevation (0-1978 m a.s.l.) and slope (0-76 • ) range. The main slope aspect varied greatly across the study area. About 43% of the total burned area is in nature conservation areas (e.g., the UE Natura 2000 network and national or regional protected areas). Generally, 30.4% (31,211 ha) of the burned area was shrubland, 19.5% (19,985 ha) was broadleaf forest, 13.1% (13,389 ha) was grassland, 9.4% (9671 ha) was conifer forest and 5.3% (5395 ha) was mixed forests (Table 3).
The RF model highlighted a root-mean-square error of 226.6 and indicated that a distance of 250 m between points effectively reduced spatial autocorrelation as the scaled semivariance at that distance was >0.7.  The best predictor in the RF model was the pre-fire land cover (CLC; Figure 7). Among land cover categories, RF model output confirmed that conifer forests are largely and positively associated with burn severity, followed, respectively, by shrubland and mixed forests (Figure 8). Higher rates of burn severity were detected at lower elevations, the second most important driver of burn severity (Figure 7) (especially from 0 to approximately 650 m a.s.l.). The models showed a positive relationship between burn severity and population density. We observed high levels of RdNBR in burned areas with more than 3000 inhabitants/km 2 . In addition, distance to railroads and roads showed a positive relationship with burn severity (Figures 8 and S2). Slope also showed a positive but weaker relationship with severity, especially increasing between 0 and 16 • ( Figure S2).

Discussion
Since many ecosystems worldwide are experiencing fire regime shifts in response to changes in climate, land use and land cover [3,5,74,75], investigations into the increasing frequency, size and severity of wildfires are becoming necessary for adopting appropriate management practices. In this work, we analyzed two target years (2007 and 2017), providing a large fire sample size for our investigation. The mean size of all the Italian fires is much lower compared to that of the fires selected, probably due to the increasing effort committed to fire suppression [76]. We considered only large wildfires since their occurrence, expansion and severity are more likely to be influenced by climatic conditions [11].
Moreover, they are expected to increase in the future given climate and land use change trends [11]. A large accumulation of fuels and the increasing fire intensity can easily drive wildfires out of control [11].
Most of the 113 fires occurred in July and August [44,51,77]. Large fires were more numerous in southern Italy, suggesting the importance of other factors such as differences in land ownership forms, regional law and regulations, and efficiency in forest prevention and fire suppression systems. In addition, recently, a relevant administrative change occurred in Italy. Until 2016, the national wildfire control system was coordinated by the National Forest Service (NFS) through their regional headquarters in synergy with the regional offices of the National Civil Protection (NCP) and their local firefighting corps. Thereafter, a newly enforced law limited the tasks of the NFS only to post-fire investigations and transferred all the fire extinction and prevention functions and the related equipment (land and air vehicles) to the NCP [27,51]. These indirect factors might have acted as noise when detecting direct relationships between spatial covariates, fire occurrence and severity.
The results of the analysis on land-cover types affected by large fires in the two target years showed that a large share of the total burned area (65%) was covered by shrubland and forests (mostly broadleaf). The co-occurrence of shrub and forest cover is a common trait of the landscape mosaics in many Mediterranean ecosystems. This is the effect of widespread secondary succession in abandoned agro-pastoral lands [15,78]. Shrubland is a very heterogeneous land-cover category in terms of species composition, structure and fuel characteristics and, together with conifer forest, is very prone to burning in the Mediterranean biome, particularly in southern Spain, France, Italy and Greece [79,80]. Generally, the shrub-forest systems burned extensively, up to 65% of the total burned area. The CLC is broadly used to assess land cover in burned areas [48,81]; however, the 25 ha of minimum mapping unit, used to describe vegetation type, could affect the precision of these estimates. Its spatial resolution and classification structure were conceived to fit a broad range of environmental EU regions and not individual countries [62].
Applied technologies and analytical methods for fire mapping and severity assessment have greatly improved in recent years [30,42,43,82]. Common indices (e.g., dNBR and RdNBR) are still broadly used in fire science [28,35,42], even though precautions in imagery selection are necessary to minimize the differences in phenology, solar angle and sensor calibration. In addition to single-image selection, composite techniques that use imagery over the entire fire season might be another robust way to compute common spectral vegetation indices [83]. In our case, combining a coarse resolution dataset (MODIS Collection 6) [34] with Landsat single-scene data made possible the detection and extraction of numerous fire records. Moreover, Landsat data guaranteed the coverage of a broad geographic area and the same spatial resolution for all the severity maps computed. We used RdNBR, a suitable spectral vegetation index for mapping fire perimeter and assessing burn severity across multiple wildfires featuring a wide range of topographic features, land cover and fire history [24]. The use of a suitable index and a spatially coherent and comprehensive largescale dataset are basic ingredients to suitably assess the environmental response to fires at the regional scale. However, RdNBR values can be influenced and biased by several factors, such as short-term vegetation recovery of non-forest land covers, fire suppression (land and airborne) and post-fire (e.g., salvage logging) management actions. Nonetheless, severity assessment is strictly related to the method applied: we used spectral changes caused by fire during the vegetative season following the event, defined as an extended assessment of remotely sensed burn severity [21]. This method accounts for first-and second-order effects caused by fire and includes delayed survivorship and mortality of vegetation [84]. Such an approach is widely used for forest and sometimes shrubland fires [85], but it can underestimate the burn severity of some land-cover types (e.g., grasslands) since a limited amount of regrowth of grasses/shrubs is common in the following vegetative season [84]. In addition, a systematic measure of burn severity in the field in many areas, despite the efforts needed, could be useful to better relate these common indices to on-the-ground fire effects [39].
Our modelling approach included land cover, anthropogenic variables and topography as possible drivers of burn severity. Its assessment in human-shaped landscapes is a very complex task because of the heterogeneity of structure and environmental constraints of landscape mosaics and of the poorly explored role of anthropogenic factors (e.g., population density and infrastructures). Our model showed that land cover was the most influential variable affecting burn severity. Several studies have highlighted the relevance of land cover, pre-fire vegetation types and greenness on burn severity in Canadian and Mediterranean forests [86,87]. Even if a metric of pre-fire vegetation cover/greenness such as NDVI can be a suitable predictor of burn severity, we used discrete data showing pre-fire types of land cover. The highest mean severity was found in conifer forests (RdNBR = 357), followed by shrublands (RdNBR = 293) and mixed forests (RdNBR = 288). It is well known that vegetation type is one of the main factors controlling burn severity computed with remotely sensed NB [40]. Conifer forests are more flammable than broadleaf forests due to their higher content of resins, terpenes and their lower foliage moisture content [88], leading to possible higher rates of burn severity, as detected by other studies in Mediterranean ecosystems [89]. Extended shrub cover can foster fire spread and burn with high severity.
Together with the amount and composition of live vegetation and dead fuel [28], topography, which directly influences fire behavior [19], can affect severity as found in several ecosystem types, as reported in the western US, China and Spain [23][24][25]90]. Our results indicate that elevation was the second most influential predictor in the model, suggesting that burn severity increased at lower elevations and secondarily on the steepest sites. Climate conditions could enhance fire ignition and spread in lower elevation areas, directly affecting burn severity. Although the topographic wetness index and the heat load index are the least important variables in the model, burn severity slightly decreased with the increasing rate of soil moisture. HLI showed no significant trend, although more exposed slopes where solar radiation is higher can affect soil moisture, fuel composition and density [91]. In Mediterranean areas, anthropogenic factors can increase forest fire risk, ignition and frequency [27,92], but their direct effects on burn severity remain unclear. Our findings suggest that population density was the third most important variable in the model, indicating a higher rate of burn severity in highly populated areas. Distance to railroads and, secondarily, distance to roads also showed a positive relationship with burn severity. A denser population could enhance the probability of ignition and frequency of wildfires [92]. Even if a higher rate of ignitions in populated areas can increase fire occurrence, if suppression systems are efficient, they could more likely reduce fire spread and burn severity. Out-of-control fires tend to be more severe in more prone-to-burn landscapes and forest types, such as conifer plantations, generally located at lower elevations, with warmer climatic conditions and where population density tends to be higher. Population density generally decreases at higher elevations where forest ecosystems and landscapes are less prone to burning and climatic conditions could reduce burn severity. Moreover, road network accessibility was found to be spatially related to human-induced fire ignitions in the US, Spain and Italy [92][93][94][95][96]. Infrastructure may affect not only fire ignitions but also their expansion and severity since fire suppression systems are largely dependent on road networks. Areas farther from infrastructure might be more affected by uncontrolled wildfires, leading to higher burn severities.
Fire expansion and severity can be reduced with specific vegetation management actions. It is well known that selective thinning and other specific silvicultural treatments can lower horizontal and vertical fuel continuity by decreasing living and dead fuel loads and modifying forest structure towards a more open canopy cover [86]. In combination with managed wildfires and prescribed fires, these actions can reduce tree density and fuel load at a landscape scale [97], mitigating the fire severity [98,99]. Such aspects, despite not directly investigated by our research, are crucial and should be evaluated for fire severity mitigation strategies.

Conclusions
Large wildfires in Mediterranean countries can cause severe environmental and socioeconomic impacts, but the analysis of their occurrence, dynamics and severity can provide relevant information for monitoring and implementing control strategies. The availability of complete and consistent records (i.e., burned area) on a national or sub-national scale is essential for understanding wildfire behavior, identifying the most influential fire drivers and implementing appropriate control strategies. The widespread forest expansion and the resultant landscape simplification triggered by the 1950-1960s overall abandonment of marginal rural areas [15] have increased the vulnerability of these areas to wildfire. Predicted climate warming and drought worsening in the Mediterranean basin, together with the increasing share of unmanaged forests and rangelands in marginal areas, can considerably raise the risk of fire ignition and consequent severity [100]. We mapped large wildfires that occurred in peninsular and insular Italy in two target years (2007 and 2017) and modelled burn severity using environmental, topographic and anthropic predictors. The selected 113 large fires have been more frequent in the southern region than in the central and northern regions. The average fire size was about 900 ha (±1240 ha) and occurred mainly in shrublands (30.4%) and broadleaf forests (19.5%). The highest severity occurred in conifer plantations and shrublands, in highly populated areas and at lower elevations. Burn severity models, at least at the landscape or regional scales, are very useful tools for pre-and post-fire forest management planning. Monitoring wildfire hazards and fire impacts on forest ecosystems is also necessary for efficient management of fire prevention and control. The final goal is to reduce the devastating effects of large forest fires on human lives, infrastructures and environmental resources. This is especially true in Mediterranean countries, where the expansion of wildland-urban interfaces (WUIs) [101] and the related increasing likelihood of dangerous fire events within their boundaries is raising concern among policymakers, land managers and the scientific community for their direct ecological and socio-ecological implications.
Supplementary Materials: The following supporting information can be downloaded at: https: //www.mdpi.com/article/10.3390/fire5060180/s1, Table S1: List of Landsat records used in the analysis; Table S2: List of locations and main features of wildfires selected for the analysis; Figure S1: Semivariogram of the random forest model residuals; Figure S2: Partial dependence plots showing relative influences of the four predictors with lower important rate of burn severity across the respective input data ranges.